Re: pg_multixact/members growing

On Wed, May 23, 2018 at 7:49 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Tiffany Thang <tiffanythang(at)gmail(dot)com> writes:
>> Our pg_multixact/members directory has been growing to more than 18GB over
>> the last couple of months. According to the documentation, the files in
>> there are used to support row locking by multiple transactions and when all
>> tables in all databases are eventually scanned by VACUUM, the older
>> multixacts are removed. In our case, the files are not removed.
>
> Hmm. What does pg_controldata tell you about NextMultiXactId,
> NextMultiOffset, oldestMultiXid, oldestMulti's DB?
> Are pg_clog/ or pg_subtrans/ or pg_multixact/offsets/ getting large?
> Is there anything at all in pg_twophase/? Is this system a replication
> master, and if so are any of its slaves lagging behind?

Some thoughts:

There are MULTIXACT_MEMBERS_PER_PAGE = 1636 members for every 8KB
page. The reported directory size implies 18GB / 8KB * 1636 =
3,859,808,256 members. Above MULTIXACT_MEMBER_SAFE_THRESHOLD =
2,147,483,647 we should be triggering emergency autovacuums to try to
reclaim space. Only ~435 million more members can be created.

Is this system now aggressively running "wraparound prevention" autovacuums?

There are MULTIXACT_OFFSETS_PER_PAGE = 2048 multixacts for every 8KB
page, so the default autovacuum_multixact_freeze_max_age should
soft-cap the size of pg_multixact/offsets at around 1.5GB ~=
400,000,000 / 2048 * 8KB.

Unfortunately autovacuum_multixact_freeze_max_age doesn't impose any
limit on the number of members. The totals can be quite explosive
with high numbers of backends, because when n backends share lock a
row we make O(n) multixacts and O(n^2) members. First we make a
multixact with 2 members, then a new one with 3 members, etc... so
that's n - 1 multixacts and (n * (n + 1)) / 2 - 1 members.

--
Thomas Munro
http://www.enterprisedb.com